Available for Research and Development Roles

Hi, I'm |

I work in the domain of

Graduate Researcher at Texas A&M building production-grade 3D human reconstruction pipelines from monocular video. Published at IEEE IGARSS and ISRO-IISc. Previously shipped ML systems at IISc and Deloitte that cut processing time by 40% and inference latency to under one second.

Download Resume View Work

Work Experience

Graduate Researcher

Texas A&M University · Advisor: Prof. Cheng Zhang
Aug 2025 - Present
  • ● 3D Scene Reconstruction: Developed an end-to-end pipeline converting monocular athlete video into dynamic 3D scenes, integrating Human3R, VGGT for pose estimation and DeepLabv3 for semantic segmentation.
  • ● Real-Time Inference: Accelerated pipeline via mixed-precision CUDA kernels and INT8 quantization, achieving 40% latency reduction for real-time biomechanical feedback during live training.
  • ● Biomechanical Analytics: Engineered a module to auto-extract metrics (jump height, center-of-mass velocity, takeoff angle) for data-driven coaching across 3+ athlete disciplines.

Project Associate

Indian Institute of Science (IISc) & ISRO Space Applications Centre
May 2023 - May 2025
  • ● Data Pipeline: Constructed an automated Python pipeline processing 100+ GB/day of HDF5 multispectral imagery, generating ~14,000 annotated frames via GDAL/Rasterio, lifting throughput by 30%.
  • ● Inference Optimization: Slashed inference latency from ~3-4 hrs to under 1 second through model pruning, INT8 quantization, and parallelized preprocessing for near-real-time operational deployment.
  • ● Experiment Tracking: Logged 200+ experiments in Weights & Biases; containerized full pipeline with Docker; scaled data workflows with Spark and Dask.
  • ● Research Impact: Co-authored peer-reviewed paper accepted at IEEE IGARSS 2025 (Brisbane) on deep-learning cyclone detection in collaboration with ISRO.

Research Intern

Georgia Institute of Technology
Jan 2023 - May 2023
  • ● Generative AI: Synthesized insider-threat datasets with GANs in PyTorch, expanding rare-class coverage and boosting out-of-distribution robustness by 15%.
  • ● Anomaly Detection: Trained unsupervised autoencoder-based detectors (+12% recall for rare events); combined with supervised ensemble (Random Forest, Logistic Regression) to reach 95% accuracy.
  • ● Data Engineering: Streamlined data-cleaning workflows with Pandas and NumPy, cutting preprocessing time by 25% across 5+ dataset splits.

Data Science Intern

Deloitte Haskins and Sells LLP
Jun 2022 - Mar 2023
  • ● Automated Auditing: Engineered Python/SQL workflows to automate journal-entry testing, slashing manual review time by 40%.
  • ● Big Data Optimization: Optimized large-scale Spark ETL pipelines, eliminating 15+ hours of weekly maintenance overhead.
  • ● Business Intelligence: Deployed 85% of Power BI dashboards and exception-analysis tools adopted across 4+ audit teams.

Education

Master of Science in Computer Science

Texas A&M University, College Station, TX
Aug 2025 - Expected May 2027

Relevant Coursework

Computer Vision and Robotic Perception 3D Computer Vision and Computer Graphics Deep Learning Deep Reinforcement Learning Analysis of Algorithms

B.Tech (Honors), Computer Science Engineering

Bennett University, Greater Noida, UP
GPA: 3.54 / 4.0 May 2023

Featured Projects

Jan 2026 – Present

Pose4DGS

  • ▹ Implemented Pose-Conditioned 4D Gaussian Splatting for generating pose-driven human avatars from monocular video using PyTorch and CUDA.
  • ▹ Engineered two zero-overhead training contributions — Pose-Space Augmentation (SMPL interpolation) and DINOv2 (ViT-B/14) feature supervision — improving unseen-pose generalization.
  • ▹ Designed a 3-tier evaluation protocol (easy/moderate/extreme) and executed a full 2×2 ablation grid on ZJU-MoCap and PeopleSnapshot.
  • ▹ Coordinated a 3-person cross-functional team; tracked 200+ runs in W&B and co-authored a 6–8 page technical report.
4D Gaussian Splatting DINOv2 SMPL W&B CUDA
Aug 2025 – Present

3D Human Reconstruction

  • ▹ Architected a neural rendering pipeline converting single-view RGB video into geometrically accurate, pose-driven 3D human meshes using NeRF and Gaussian Splatting.
  • ▹ Modeled human motion interactions by combining neural implicit representations with SMPL body priors, targeting sub-centimeter surface accuracy.
  • ▹ Evaluated reconstruction quality against SOTA baselines using PSNR, SSIM, LPIPS on ZJU-MoCap and Human3.6M benchmarks.
  • ▹ Iterated on deformation-network architectures and loss functions to improve cross-subject generalization under Prof. Cheng Zhang.
NeRF Gaussian Splatting SMPL COLMAP PyTorch
Aug – Dec 2025

Recurrent Attention Model

  • ▹ Reproduced RAM in PyTorch for MNIST classification using sequential local glimpses trained with REINFORCE policy gradients, reaching 96.64% accuracy with only 6 glimpses (~49% of pixels).
  • ▹ Deployed a learned value baseline to reduce gradient variance; ablated glimpse count (4/6/8), patch resolution, and reward shaping strategies.
  • ▹ Visualized attention trajectories confirming systematic digit-edge fixation patterns — matching full-image CNN baselines at a fraction of the compute.
  • ▹ Released a reproducible open-source codebase with scripted evaluations, training-stability analysis, and result visualizations.
REINFORCE PyTorch Attention NumPy
May 2023 – May 2025

CycloAI

  • ▹ Co-built an end-to-end multi-sensor perception pipeline processing 100+ GB/day of HDF5 multispectral satellite data with a two-stage detector (Center-Locator + Intensity-Estimator).
  • ▹ Slashed inference from ~3–4 hrs to <1 sec via model pruning, INT8 quantization, and parallelized preprocessing — enabling near-real-time ISRO operational deployment.
  • ▹ Boosted minority-class recall by 12% using focal loss and SMOTE; assessed detection across 5+ spectral bands (mAP, RMSE); monitored 200+ runs in W&B.
  • ▹ Published at IEEE IGARSS 2025 (Brisbane) — containerized full pipeline with Docker for reproducible deployment.
PyTorch CUDA Docker Spark GDAL
Aug – Dec 2022

CNN Architectures from Scratch

  • ▹ Coded ResNet and VGG variants from first principles in PyTorch without pre-trained weights, building deep intuition for residual connections and feature hierarchies.
  • ▹ Assessed convergence and generalization on CIFAR-10 and ImageNet subsets with gradient-flow analysis and scripted benchmarks.
  • ▹ Diagnosed vanishing gradient issues in deep VGG, demonstrating ~2× faster convergence with skip connections in ResNet counterparts.
ResNet VGG PyTorch NumPy
May 2021 – May 2022

Chest X-Ray Pathology Classification

  • ▹ Fine-tuned DenseNet-121 on CheXpert, matching published AUC benchmarks for 14 thoracic pathologies in a multi-label classification setting.
  • ▹ Applied SMOTE and focal loss to address severe class imbalance, raising minority-class recall by 8%.
  • ▹ Built an end-to-end evaluation workflow with automated ROC/AUC computation and model versioning for reproducible experiments.
DenseNet CheXpert Medical AI Scikit-Learn

Publications

Comparative Analysis of Deep Learning Models for Cyclone Detection

IEEE International Geosciences and Remote Sensing Symposium (IGARSS) 2025

A Visual-Analytical Approach for Automatic Detection of Cyclonic Events in Satellite Observations

40th Annual Symposium on Space Science and Technology, ISRO-IISC (2024)

Generative Adversarial Network Based Approach towards Synthetically Generating Insider Threat Scenarios

Journal of Software Engineering and Applications (2023)

Technical Skills

Languages & Frameworks

Python C/C++ SQL R PyTorch TensorFlow Scikit-Learn NumPy/Pandas

Perception & Computer Vision

3D Reconstruction 4D Gaussian Splatting Pose Estimation Object Detection Segmentation SMPL/SMPLx DINOv2

Tools & Infrastructure

CUDA Docker AWS Spark/Dask Weights & Biases COLMAP GDAL/Rasterio Git

Evaluation & Methods

PSNR / SSIM / LPIPS mAP / RMSE AUC-ROC Model Pruning & Quantization Ablation Study Design

Let's Build Something

Based in College Station, TX. Actively seeking Summer 2026 research internships and full-time opportunities in Computer Vision, 3D Reconstruction, and Applied ML. Let's talk.

mayesh@tamu.edu LinkedIn Full Resume (2-page)
© 2026 Mayesh Mohapatra.